Inference FPS Fixed on Hailo 8L with Different YOLOv8 Models(batch size = 1)

General
1

Hi all,

I have successfully set up my pipeline using customized data and tested it across different model sizes: YOLOv8n, YOLOv8s, and YOLOv8m.

When testing the performance from a single-channel stream to multi-channel streams, I observed the following behavior:

  • The inference FPS displayed in hailortcli monitor is fixed at 4 FPS for all models.
  • However, the total FPS (as displayed in tools like fpsdisplaysink) remains at around 19-20 FPS.
  • The CPU utilization varies depending on the model size, but the inference FPS does not change.

It seems the inference frame rate is being locked at a certain value, regardless of the model size or input configuration.

I’m wondering if this is due to:

  1. Internal HailoRT limitations or specific settings.
  2. Queue or buffer configurations in my GStreamer pipeline.
  3. Resource constraints or an implicit frame rate cap.

Has anyone encountered similar behavior? Any suggestions or insights on where to look to resolve this would be greatly appreciated.